4 ms·
Is is possible that although neural nets can overfit as this paper shows, practitioners just stop training early before this happens? And/or they use a validati
by miles7 10y ago
Is is possible that although neural nets can overfit as this paper shows, practitioners just stop training early before this happens? And/or they use a validation set? Would that be enough to explain the good generalization despite the huge number of parameters
- argonaut 10y agoIt doesn't explain why it learns anything in the first place (e.g. why it doesn't just overfit from the start; or learn some weak signals, then start overfitting).